CN-121997051-A - Line spectrum characteristic migration underwater sound signal generation method based on large model physical constraint

CN121997051ACN 121997051 ACN121997051 ACN 121997051ACN-121997051-A

Abstract

The invention relates to a line spectrum characteristic migration underwater sound signal generation method based on large model physical constraint, which comprises the steps of constructing an underwater sound target signal data set, preprocessing the underwater sound target signal data set, inputting the preprocessed data set into a physical enhancement type LSFT-GAN generation model for model training to obtain an underwater sound signal generation model, wherein the physical enhancement type LSFT-GAN generation model integration generator, a line spectrum characteristic encoder, a large language model reasoning module and a physical-characteristic alignment module, and generating a trusted underwater sound target signal conforming to a physical rule by utilizing the underwater sound signal generation model. The invention not only maintains the advantages of LSFT-GAN in multi-class feature migration, but also ensures that the generated signal accords with the underwater sound physical propagation rule by utilizing LLM, thereby remarkably improving the authenticity and the credibility of the generated signal and effectively solving the problem that a data driving model can generate a 'physical violation' sample.

Inventors

LI RONGSHENG
WU YANXIA
SHI CHANGTING
LIN DAN
CHEN ZUDONG
CHEN ZHI
LIU SHUYONG
FU YAN

Assignees

哈尔滨工程大学

Dates

Publication Date: 20260508
Application Date: 20260212

Claims (10)

1. A line spectrum characteristic migration underwater sound signal generation method based on large model physical constraint is characterized by comprising the following steps: Constructing an underwater sound target signal data set, wherein the underwater sound target signal data set comprises original signals of various ship categories and environmental noise; preprocessing the underwater sound target signal data set; Inputting the preprocessed data set into a physical enhanced LSFT-GAN generation model for model training to obtain an underwater sound signal generation model, wherein the physical enhanced LSFT-GAN generation model is integrated with a generator, a line spectrum feature encoder, a large language model reasoning module and a physical-feature alignment module; And generating a trusted underwater sound target signal conforming to a physical rule by using the underwater sound signal generation model.
2. The method of generating a line spectral feature migration underwater sound signal based on large model physical constraints of claim 1, wherein preprocessing the underwater sound target signal data set comprises: Denoising, framing and short-time Fourier transformation are carried out on the original signal, and Z-Score standardization processing is adopted to eliminate energy level differences under different sea conditions and unify data distribution.
3. The method for generating a line spectrum feature migration underwater acoustic signal based on large model physical constraints of claim 1, wherein inputting the preprocessed data set into the physical enhancement type LSFT-GAN generation model comprises: evaluating the underwater sound target signal data set to generate dynamic physical feedback information; splicing a preset static structured prompt word template with the dynamic physical feedback information to obtain a composite text description; inputting the composite text description into the large language model reasoning module, and carrying out multi-step logical reasoning based on the text description to obtain a physical constraint vector; extracting a prototype line spectrum feature code which is randomly extracted from the preprocessed data set and serves as a reference sample of a target generation category by using the spectrum feature encoder; Inputting the physical constraint vector and the prototype line spectrum feature code into the physical-feature alignment module for multi-mode fusion and correction to obtain a physical consistency line spectrum feature code; And injecting the physical consistency line spectrum feature codes into a decoder level of a generator by using an adaptive instance normalization mechanism, and controlling the line spectrum morphology and physical properties of the generated signals.
4. The method for generating a line spectrum feature migration underwater sound signal based on large model physical constraints according to claim 3, wherein the large language model reasoning module performs multi-step logical reasoning based on text description, and the obtaining of the physical constraint vector comprises: Constructing a structured prompting word template, wherein the structured prompting word template comprises target type, fine motion state, sea state level, diving depth and receiving and transmitting distance parameters; The large language model reasoning module activates a thinking chain mechanism to deduce in stages according to input text description, firstly deduces a sound source mechanism, estimates fundamental frequency and leaf frequency positions, secondly deduces a propagation channel, calculates high-frequency attenuation slope and Doppler frequency shift quantity, finally deduces an environmental effect, judges whether multipath interference fringes exist or not, and maps a deduction result into a numeric physical constraint vector.
5. The method for generating a line spectrum feature migration underwater sound signal based on large model physical constraints of claim 4, wherein activating the mind chain mechanism comprises the steps of: estimating the total sound source level of the radiation noise according to the target type and the navigational speed, and deducing the fundamental frequency and the leaf frequency position; Propagation channel deduction, namely judging a propagation model according to the receiving and transmitting distance and sea depth, calculating the attenuation slope of a high-frequency component, and calculating the frequency shift direction and magnitude of a spectral line according to the relative radial speed by utilizing a Doppler formula; the environmental effect deduction is that the reflection coefficient of the seabed substrate and the wave scattering of the sea surface are synthesized, and whether multipath interference fringes exist or not is deduced; And the physical constraint vector coding is used for mapping the deduction results of the sound source mechanism, the propagation channel and the environmental effect into a numerical physical constraint vector, wherein the components of the physical constraint vector at least comprise normalized fundamental frequency, energy attenuation coefficient, harmonic richness coefficient, environmental noise ground color coefficient and Doppler frequency shift factor.
6. The method for generating a line spectrum feature migration underwater sound signal based on large model physical constraint according to claim 3, wherein the physical-feature alignment module performs multi-mode fusion, and an adaptive instance normalization mechanism is adopted: And correcting the physical-characteristic alignment module by adopting a residual error connection mechanism.
7. A method of generating a line spectral feature migration underwater acoustic signal based on large model physical constraints as claimed in claim 3, wherein injecting the physical consistency line spectral feature code into the decoder hierarchy of a generator comprises: generating affine transformation parameters by using the physical consistency line spectrum feature codes Bias factor ; Generating a physical mask through a gating network using the physical constraint vector; Based on the physical mask and scaling factor Bias factor And performing feature modulation, wherein the physical mask is used for dynamically suppressing artifact noise which is inconsistent with the physical scene in the spatial dimension of the feature map.
8. The method for generating a line spectrum feature migration underwater sound signal based on large model physical constraints of claim 7, wherein performing feature modulation comprises: Wherein, the For the element-by-element multiplication, In the case of a physical mask, In order for the scaling factor to be a factor, As a result of the bias factor, In order to output the characteristics of the feature, In order to input the characteristics of the feature, Is a normalization operation.
9. A method of generating a line spectrum feature migration underwater acoustic signal based on large model physical constraints as claimed in claim 3, wherein each layer of convolution of the decoder of the generator incorporates a physical constraint gating unit; The working mechanism of the physical constraint gating unit is as follows: Wherein, the The function is activated for Sigmoid, For the element-by-element multiplication, In the form of a physical constraint vector, As a matrix of weights, the weight matrix, In order to output the characteristics of the feature, As a result of the bias term, The operation is normalized for the adaptive instance.
10. The method for generating a line spectrum feature migration underwater sound signal based on large model physical constraints of claim 1, wherein performing model training comprises: Constructing a mixed objective function comprising a countermeasure loss, a loop consistency loss, a line spectrum reconstruction loss and a physical consistency loss, and performing end-to-end joint training on the model.

Description

Line spectrum characteristic migration underwater sound signal generation method based on large model physical constraint Technical Field The invention relates to the technical fields of underwater acoustic signal processing, deep learning, large model application and calculation of marine acoustics, in particular to a line spectrum characteristic migration underwater acoustic signal generation method based on large model physical constraint. Background The high-quality and accurate-labeling underwater sound target signal data is the basis for training a high-performance underwater sound target recognition model. However, practical marine testing is costly, data acquisition is extremely difficult, resulting in serious class imbalance and insufficient scene coverage problems for existing data sets. For example, data for extreme maneuver states of non-cooperative targets or under certain complex sea conditions, tend to be blank. Aiming at the problem of data scarcity, the existing generation methods are mainly divided into two types: And one is a simulation method based on a physical model. The method is based on strict wave equation solution, the generated signal physical rule is strict, and the propagation loss calculation is accurate. However, the physical model parameter settings are extremely complex and often based on idealized assumptions, it is difficult to simulate the complex randomness of a real marine environment, resulting in a generated signal that is too "clean" or "mechanical", lacks "realism", and has poor generalization ability on recognition models. And secondly, a data driving method based on deep learning. It generates new samples by learning the probability distribution of large amounts of real data. The generated signal texture has high fidelity and the auditory sense is closer to the real sound recording. However, because the model is essentially a "black box" that fits the data distribution, lacking an understanding of the acoustic physics mechanism, existing GAN models tend to produce "statistically reasonable but physically impossible" samples. For example, the Doppler frequency shift direction in the generated signal is opposite to the motion direction, the propagation loss does not accord with the distance attenuation rule, the abnormity that the energy of the remote signal is enhanced instead occurs, or the line spectrum harmonic structure is disordered and does not accord with the frequency multiplication relation of mechanical operation. If the physical violation samples are mixed into the training set, the recognition model can be seriously misled to learn the wrong characteristic mode, and the reliability of the recognition model in actual sea trial is reduced. Although the existing LSFT-GAN method realizes the migration of line spectrum characteristics, the characteristic control mainly depends on the image characteristics of a reference sample, and the logic level adjustment cannot be performed according to specific working condition description. The large language model has strong knowledge retrieval, logical reasoning and little sample learning capability, and can understand physical concepts and conduct parameter deduction. Therefore, LLM is introduced into the field of underwater acoustic signal generation, physical knowledge is used as a priori constraint generation process, and the method is a key way for realizing the fusion of data driving and knowledge driving and improving the reliability of generated signals. Disclosure of Invention In order to solve the technical problems, the invention provides a line spectrum characteristic migration underwater sound signal generation method based on large model physical constraint, which aims to integrate the texture authenticity generated by data driving and the physical correctness of knowledge driving reasoning and improve the physical credibility of the generated signals and the effectiveness in practical application. In order to achieve the above object, the present invention provides the following solutions: A line spectrum characteristic migration underwater sound signal generation method based on large model physical constraint comprises the following steps: Constructing an underwater sound target signal data set, wherein the underwater sound target signal data set comprises original signals of various ship categories and environmental noise; preprocessing the underwater sound target signal data set; Inputting the preprocessed data set into a physical enhanced LSFT-GAN generation model for model training to obtain an underwater sound signal generation model, wherein the physical enhanced LSFT-GAN generation model is integrated with a generator, a line spectrum feature encoder, a large language model reasoning module and a physical-feature alignment module; And generating a trusted underwater sound target signal conforming to a physical rule by using the underwater sound signal generation model. Optionally, preprocessi